Learning a goal-oriented model for energy efficient adaptive applications in data centers

Abstract

This work has been motivated by the growing demand of energy coming from the Information Technology (IT) sector. We propose a goal-oriented approach where the state of the system is assessed using a set of indicators.These indicators are evaluated against thresholds that are used as goals of our system. We propose a self-adaptive context-aware framework, where we learn both the relations existing between the indicators and the eect of the available actions over the indicators state. The system is also able to respond to changes in the environment, keeping these relations updated to the current situation. Results have shown that the proposed methodology is able to create a network of relations between indicators and to propose an eective set of repair actions to contrast suboptimal states of the data center. The proposed framework is an important tool for assisting the system administrator in the management of a data center oriented towards Energy Eciency (EE), showing him the connections occurring between the sometimes contrasting goals of the system and suggesting the most likely successful repair action(s) to improve the system state, both in terms of EE and Quality of Service (QoS).

Toward sustainable data centers: a comprehensive energy management strategy

“…The problem can be also solved as a single-objective problem when all but one of the objectives have a target value, as a constraint can be placed on those objectives. When a target value can be identified for all the objectives, the problem can be solved by means of goal-oriented adaptation, which is able to learn from monitored data the impact of repair actions on the value of the goals and apply the most convenient action when any of them deviates from its target value [70]. Another possibility for solving true multi-objective problems consists of computing all or a representative set of Pareto optimal solutions (those in which it is impossible to make any objective better off without making at least another one worse off), which are usually derived by means of evolutionary algorithms [74,32].…”

“…Those high-level objectives are commonly referred as Business Level Objectives (BLO). The Green Grid Data Center Maturity Model [9], which suggests best practices for energy efficient data centers, can be also used as reference to define constraints [70]. The energy-aware manager configures the management controllers in order to fulfill the BLOs.…”

“…According to this, the energyaware manager must implement a self-adaptive loop that monitors the system status, evaluates the BLOs fulfillment under those conditions, and aligns the configuration of the management controllers to ensure the fulfillment of the BLOs [62]. The effect of different configurations over the BLOs can be learned from available data using machine learning techniques [70].…”

Optimizing Monitorability of Multi-cloud Applications

“…-Monitored dimensions which are not directly measured but their trends are estimated exploiting the existing dependencies among metrics [18]. A Bayesian Network is adopted to express the likelihood of a metric to increase or decrease its value when the value of another metric increases or decreases.…”

Data Movement in the Internet of Things Domain

“…Since it is difficult in the general case to design at run time all the possible relationships between indicators, learning tools, such as Bayesian Networks, can be used to derive the relations among the indicators and the goals, and to evaluate the effect of selecting a mode to provide data on these indicators [14].…”